Stream: interviews

Topic: 629: Programming with LLMs

Logbot (Feb 19 2025 at 19:00):

For the past year, David Crawshaw has intentionally sought ways to use LLMs while programming, in order to learn about them. He now regularly use LLMs while working and considers their benefits a net-positive on his productivity. David wrote down his experience, which we found both practical and insightful. Hopefully you will too! :link: https://changelog.fm/629

Nabeel S (Feb 20 2025 at 15:00):

Are LLMs like running shoes for a runner, or more like a Segway? :joy::joy:

Ron Waldon-Howe (Feb 21 2025 at 00:26):

I still can't get past the plagiarism and unfairness
Why isn't the USA DoJ hounding Altman and Zuckerberg to the degree they bullied poor Aaron Schwartz, who downloaded virtually nothing by compariso

Ron Waldon-Howe (Feb 21 2025 at 00:28):

Models built strictly from public domain and/or no-restrictions licenses like BSD are fine
Anyone peddling anything else should be facing lawsuits or jail time, in a just world

Notification Bot (Feb 22 2025 at 06:22):

A message was moved here from #interviews > test general chat by Adam Stacoviak.

Adam Stacoviak (Feb 22 2025 at 06:22):

@Tim Uckun not sure if this belongs here but I made an assumption.

valon-loshaj (Feb 22 2025 at 18:45):

great conversation, found myself nodding in agreement with a lot of the observations david mentioned during the interview. i’ve ground a lot of the same observations over the past 2 years playing with these tools.

one thing i didn’t hear mentioned too much was agent swarming. check out tools like “swarm” and “crew-ai” that allow you to have multiple models handle a single request.

the outputs from the “crew” of models can be used to come up with a consensus response based on the collective outputs.

i’ve found that using the models like this gives some really good results.

Ron Waldon-Howe (Feb 22 2025 at 19:49):

swarm/crew reminds me of a podcast discussion (maybe Practical AI, or Oxide & Friends) where they had multiple models simulate a focus group, with each model adopting a specific persona

valon-loshaj (Feb 23 2025 at 13:47):

that’s a practical ai episode, i’m still looking for it again lol.

that was when i first heard about it.

if anyone remembers that episode, please let me know!

John Johnson (Feb 26 2025 at 22:04):

I agree that prompt engineering tricks are rapidly changing, and that a better approach is to treat prompting as if you're speaking to an intelligent person lacking sufficient context.

Tim Uckun (Feb 27 2025 at 00:33):

I can’t wait for ai to tell me to buy a Samsung phone

Ron Waldon-Howe (Feb 27 2025 at 00:59):

I just came across this which is bizarre: https://arstechnica.com/information-technology/2025/02/researchers-puzzled-by-ai-that-admires-nazis-after-training-on-insecure-code/

Tim Uckun (Feb 27 2025 at 01:21):

I guess those ideas are somewhere in the training data so they come out sooner or later.

Ron Waldon-Howe (Feb 27 2025 at 01:32):

Maybe training these models on Reddit and 4chan was not such a good idea...

Tim Uckun (Feb 27 2025 at 03:32):

or facebook or xitter or Joe Rogan podcast transcripts

Sukhdeep Brar (Feb 27 2025 at 13:22):

"onboarding" the AI _each_ time you talk to it is a pain. I still need to find a better system of building local "onboarding docs", and passing them into various AI tools.
Otherwise, all the things that make onboarding a junior dev still apply.

Tim Uckun (Feb 27 2025 at 21:48):

English is a terrible language to use when specifying what you want your program to to do.

Ron Waldon-Howe (Feb 28 2025 at 00:00):

English is a terrible language
Fixed it for you :)

Jamie Tanna (Mar 01 2025 at 13:14):

Did we ever work out what the "local Claude alternative product" was? (not yet finished, but don't think it's in the show notes)

Last updated: Apr 04 2025 at 01:15 UTC